Measure Twice

12 rules that stop your AI from rushing, guessing, and making you repeat yourself.

Install
cmdop skills install agensi-measure-twice

Measure Twice is a skill published by agensi that installs 12 behavioral rules into an AI agent to reduce a specific class of repeated mistakes. Each rule was derived from a real session failure rather than theoretical best practices. The rules change how an agent approaches every task it receives.

The core loop the skill enforces is: quote the user’s exact words before acting, read the current state of any file before editing it, write a single sentence describing the intended change in the form “Currently X, want Y, changing Z”, implement the change, then re-read the file to confirm the result matches what was requested. This verification step replaces the common pattern of an agent declaring a task complete without checking.

The rules specifically address the following failure modes: editing files without reading their current contents first, paraphrasing a user’s request and dropping the detail that mattered, retrying the same broken fix with only minor variation, placing changes in the wrong file or wrong scope, breaking adjacent code while fixing a targeted issue, and losing track of the original request after several exchanges.

Measure Twice is a skill with no tools and no external service connections. It carries no package version, license, or repository. It is appropriate when the primary problem is agent carelessness or loop failures in coding workflows, not when additional integrations or data access are needed.

Use cases

  • Prevent an agent from editing a file without first reading its current contents
  • Stop an agent from paraphrasing a request and losing specific details the user stated
  • Enforce a verification step so the agent re-reads a file after editing instead of assuming success
  • Break a loop where the agent retries the same failing fix with slightly different values
  • Ensure the agent quotes the user's exact words before acting on a task
  • Keep the agent scoped to the correct file and avoid unintended changes to adjacent code

When to use it

  • The agent repeatedly claims a task is done without actually verifying the result
  • The agent edits files without reading their current state first
  • The agent keeps retrying the same approach that has already failed
  • The agent paraphrases requests and loses important specifics the user provided
  • Coding sessions involve multiple back-and-forth exchanges where the original goal gets lost

When not to use it

  • The task requires the agent to call external APIs or databases — this skill provides no tools or integrations
  • A package manager or versioned install is required — no package identifier or registry is available for this skill
  • The problem is capability gaps rather than behavioral carelessness
  • The agent platform does not support skill-based instruction injection